METR chief scientist Ryan Greenblatt denies OpenAI accusations over fired safety researchers and leaked memos
Based on this article, OpenAI has made false accusations about me. The three researchers fired from OpenAI were told I facilitated architecture leaks to The Information. This is false. A journalist from The Information did reach out to me shortly before that article came out. I did not comment at all beyond asking the journalist to make it clear that a blog post I wrote which was referenced in the article was preexisting and public. The article also mentions an internal OpenAI monitorability proposal. I have heard that this memo contained highly sensitive architecture information, and that OpenAI accused me of accessing and reading it. This accusation is also false. As reported in the article, I viewed a version of the proposal that did not have sensitive IP and was approved for sharing with me. I believe OpenAI now knows this accusation was based on a conflation of these two documents, though I do not know if they have clearly retracted that accusation. I do not and have never solicited proprietary information, and the three employees who were recently fired from OpenAI never leaked IP to me. As the article mentions, I did have Signal chats with them that had disappearing messages on; they’re professional contacts that I’ve had on Signal for a while, and use of Signal is common in the tech industry. During the Hugging Face investigation, I was careful to avoid sending Signal messages with confidential information, though I did sometimes discuss high level things on Signal calls. As far as I understood, this was normal.
Context
The public dispute between OpenAI and an external safety evaluator underscores growing friction over chain of thought monitorability, internal architecture secrecy, and informal communication channels between frontier lab staff and outside auditors.
On October 22, The Wall Street Journal reported that OpenAI fired safety researchers Tomek Korbak, Mikita Balesni, and Jasmine Wang in September after security officers inspected their unlocked personal phones and found disappearing Signal messages with external safety researcher Ryan Greenblatt. The three researchers had advocated inside OpenAI for preserving human readable chain of thought reasoning rather than moving to opaque latent reasoning architectures.
Deep dive
Responding on October 23 to a Wall Street Journal investigation by Maxwell Zeff, Keach Hagey, and Berber Jin, METR chief scientist and former Redwood Research chief scientist Ryan Greenblatt stated that OpenAI had made false accusations about him. According to the Journal, OpenAI fired safety researchers Tomek Korbak, Mikita Balesni, and Jasmine Wang last month after security officers asked to inspect their personal phones and found disappearing Signal messages in chats with Greenblatt during an inquiry into leaks about OpenAI's Astra model architecture.
In his post, Greenblatt wrote that the three fired OpenAI researchers did not help him leak information to The Information and that he has never leaked information to The Information or assisted anyone in doing so. Greenblatt also denied OpenAI's claim that he read a sensitive internal OpenAI memo, stating that he neither saw nor read the document and only learned of its existence after OpenAI accused him.
According to the Wall Street Journal report, Korbak, Balesni, and Wang had spent the summer campaigning inside OpenAI to keep AI chain of thought reasoning visible to human monitors rather than adopting new architectures that reason silently in continuous latent representations. OpenAI told the Journal that the trio were terminated for violating clear policies on handling sensitive information and not for raising safety concerns, while affirming that preserving frontier model monitorability requires an industrywide commitment.
What to expect next
As frontier labs tighten internal information security to protect novel reasoning architectures and proprietary training pipelines, independent evaluators like METR and Redwood Research face a narrower boundary between sanctioned pre deployment auditing and informal researcher collaboration across organizations, particularly when new latent reasoning designs reduce human visibility into how frontier models arrive at complex actions during external safety and alignment testing.